Meta Just Gave AI Perfect Memory?! (CLM's)

If you have ever run a long session in Claude Code or Codex, you know the pain: halfway through, the agent silently forgets a crucial rule you set at the beginning. That happens because an external software harness—not the AI model—decides when to summarize and what to delete. 🔔 Subscribe:    / @cloud-codes   💙 Become a Member:    / @cloud-codes   🐦 Twitter/X: https://x.com/cloud_codes 💬 Discord:   / discord   In this breakdown, Cloud Codes explores "Context Language Models" (CLMs)—breakthrough research from Meta Superintelligence Labs, the University of Washington, MIT, and Trillium Labs: • The Paper Roll vs. The File: Why current append-only context windows force crude summarization, while CLMs treat context as an open file the model edits via Bash. • Emerging Self-Directed Behaviors: How models autonomously wrote loops to purge useless web search results, created a "notes" role, and updated an in-context scoreboard 163 times while keeping token length pinned at 6 to 8K. • Zero-Shot Efficiency: Outperforming Codex-style summarization on BrowseComp-Plus (59.4% vs 53.4%) and scoring 5% higher on 12-hour EdgeBench runs while using 59% less compute. • The RL Benchmark Twist: Why the headline 47.6% RL gain on Qwen3.5-9B was actually an accuracy tie against a trained summary harness (42.5% vs 42.1%)—and why the real win was 38.8% lower compute. • The Caching Bottleneck: Why editing the middle of a prompt destroys standard prefix caching, and how Suffix Cache Reuse cuts SGLang server compute by 35%. • The Unsolved Security Risk: How editable context allows prompt injections and unauthorized model-written instructions to persist indefinitely across turns. 🔗 Verified Sources: • Context Language Models Paper (arXiv:2609.37725): https://arxiv.org/abs/2609.37725 • Full Paper HTML: https://arxiv.org/html/2609.37725v1 • Official Code Repository: https://github.com/facebookresearch/c... • Anthropic Claude Code Context Compaction Docs: https://code.claude.com/docs/en/how-c... • Alex Zhang (RLM Author) Analysis on X: https://x.com/a1zhang/status/21054099... ⏱️ Chapters: 0:00 - Why AI Agents Forget Early Rules 0:46 - The Problem With Claude Code & Codex Compaction 1:29 - Not an RLM: How CLMs Actually Differ 2:02 - The Paper Roll vs. The Editable Context File 2:57 - What AI Does When Given Full File Access 3:22 - The 163-Edit In-Context Scoreboard 3:49 - Zero-Shot Benchmarks: Cutting Compute by 59% 5:13 - The +47.6% RL Claim (And Table 2's Twist) 6:05 - The Server Bottleneck: Breaking Prefix Caching 6:32 - Suffix Cache Reuse in SGLang 7:08 - Open-Source Code, Non-Commercial License 7:36 - The Dangerous Safety Catch (Persistent Injections) 8:13 - Final Verdict: Who Should Hold the Scissors? #meta #ai #machinelearning #clm #claudecode #codex #llm #cloudcodes User Queries: why does claude code forget instructions how does context language models clm work meta superintelligence labs context language models difference between rlm and clm ai claude code compact turns drops rules how to fix ai agent memory loss in long sessions suffix cache reuse sglang prefix caching qwen 3.5 9b browsecomp plus benchmark can an ai model edit its own context window context language models arxiv 2609.37725 prompt injection persistent editable context clm